feat: add per-model concurrency limits

offline-backend-20260512/backend/app/gateway/routers/admin_active_runs.py

@@ -54,6 +54,7 @@ class ActiveRunItem(BaseModel):
     user_name: str = ""
     agent_id: str = ""
     agent_name: str = ""
+    model_name: str = ""
     thread_title: str = ""
     kind: str = "对话"
     multitask_strategy: str = "reject"
@@ -245,6 +246,7 @@ async def list_active_runs(request: Request) -> ActiveRunsResponse:
                 user_name=user_emails.get(user_id) or user_id or "未知用户",
                 agent_id=assistant_id,
                 agent_name=agent_name,
+                model_name=str((r.metadata or {}).get("model_name") or ""),
                 thread_title=(meta_row.display_name if meta_row else "") or "",
                 kind=_kind_for(metadata),
                 multitask_strategy=r.multitask_strategy,

offline-backend-20260512/backend/app/gateway/routers/system_settings.py

@@ -13,6 +13,7 @@ from app.gateway.deps import get_optional_user_from_request
 from deerflow.config import get_app_config
 from deerflow.config.system_settings import (
     LeaderboardSettings,
+    ModelConcurrencySettings,
     PromptPrefixSettings,
     PublishSquare,
     PublishSquaresSettings,
@@ -128,6 +129,18 @@ class SkillCompressionUpdateRequest(BaseModel):
     keep_index_in_prompt: bool | None = None
 
 
+class ModelConcurrencyResponse(BaseModel):
+    enabled: bool = Field(..., description="Whether per-model concurrency limiting is active")
+    default_limit: int = Field(..., description="Fallback max concurrent runs for models without an explicit override")
+    model_limits: dict[str, int] = Field(default_factory=dict, description="Per-model overrides: model_name -> max concurrent runs")
+
+
+class ModelConcurrencyUpdateRequest(BaseModel):
+    enabled: bool | None = None
+    default_limit: int | None = Field(default=None, ge=1, le=10_000)
+    model_limits: dict[str, int] | None = None
+
+
 def _skill_compression_to_response(settings: SkillCompressionSettings) -> SkillCompressionResponse:
     return SkillCompressionResponse(
         enabled=settings.enabled,
@@ -136,6 +149,26 @@ def _skill_compression_to_response(settings: SkillCompressionSettings) -> SkillC
     )
 
 
+def _model_concurrency_to_response(settings: ModelConcurrencySettings) -> ModelConcurrencyResponse:
+    cleaned: dict[str, int] = {}
+    for name, limit in (settings.model_limits or {}).items():
+        model_name = str(name).strip()
+        if not model_name:
+            continue
+        try:
+            normalized_limit = int(limit)
+        except (TypeError, ValueError):
+            continue
+        if normalized_limit < 1:
+            continue
+        cleaned[model_name] = min(normalized_limit, 10_000)
+    return ModelConcurrencyResponse(
+        enabled=settings.enabled,
+        default_limit=max(1, min(int(settings.default_limit), 10_000)),
+        model_limits=cleaned,
+    )
+
+
 @router.get("/skill-compression", response_model=SkillCompressionResponse)
 async def get_skill_compression() -> SkillCompressionResponse:
     """Return the skill-compression configuration (readable by any user).
@@ -174,6 +207,47 @@ async def update_skill_compression(body: SkillCompressionUpdateRequest, request:
     return _skill_compression_to_response(saved.skill_compression)
 
 
+@router.get("/model-concurrency", response_model=ModelConcurrencyResponse)
+async def get_model_concurrency(request: Request) -> ModelConcurrencyResponse:
+    """Return the per-model concurrency-limit configuration. Admin only."""
+    await _require_admin(request)
+    return _model_concurrency_to_response(load_system_settings().model_concurrency)
+
+
+@router.put("/model-concurrency", response_model=ModelConcurrencyResponse)
+async def update_model_concurrency(
+    body: ModelConcurrencyUpdateRequest,
+    request: Request,
+) -> ModelConcurrencyResponse:
+    """Update the per-model concurrency-limit configuration. Admin only."""
+    await _require_admin(request)
+
+    current = load_system_settings()
+    fields = body.model_dump(exclude_unset=True)
+    update: dict[str, object] = {}
+    if "enabled" in fields and fields["enabled"] is not None:
+        update["enabled"] = bool(fields["enabled"])
+    if "default_limit" in fields and fields["default_limit"] is not None:
+        update["default_limit"] = max(1, min(int(fields["default_limit"]), 10_000))
+    if "model_limits" in fields and fields["model_limits"] is not None:
+        cleaned: dict[str, int] = {}
+        for name, limit in (body.model_limits or {}).items():
+            model_name = str(name).strip()
+            if not model_name:
+                continue
+            try:
+                normalized_limit = int(limit)
+            except (TypeError, ValueError):
+                continue
+            if normalized_limit < 1:
+                continue
+            cleaned[model_name] = min(normalized_limit, 10_000)
+        update["model_limits"] = cleaned
+    settings = current.model_concurrency.model_copy(update=update)
+    saved = save_system_settings(current.model_copy(update={"model_concurrency": settings}))
+    return _model_concurrency_to_response(saved.model_concurrency)
+
+
 def _leaderboard_to_response(settings: LeaderboardSettings) -> LeaderboardSettingsResponse:
     return LeaderboardSettingsResponse(
         exclude_user_ids=list(dict.fromkeys([uid for uid in settings.exclude_user_ids if uid])),

offline-backend-20260512/backend/app/gateway/services.py

@@ -15,13 +15,15 @@ import time
 from collections.abc import Mapping
 from typing import Any
 
-from fastapi import HTTPException, Request
-from langchain_core.messages import HumanMessage
-
-from app.gateway.deps import get_run_context, get_run_manager, get_stream_bridge
-from app.gateway.utils import sanitize_log_param
-from deerflow.config.system_settings import load_system_settings
-from deerflow.runtime import (
+from fastapi import HTTPException, Request
+from langchain_core.messages import HumanMessage
+
+from app.gateway.deps import get_run_context, get_run_manager, get_stream_bridge
+from app.gateway.utils import sanitize_log_param
+from deerflow.config import get_app_config
+from deerflow.config.agents_config import load_agent_config
+from deerflow.config.system_settings import load_system_settings
+from deerflow.runtime import (
     END_SENTINEL,
     HEARTBEAT_SENTINEL,
     ConflictError,
@@ -366,7 +368,7 @@ def build_run_config(
     return config
 
 
-def _collect_requested_agent_ids(body: Any) -> set[str]:
+def _collect_requested_agent_ids(body: Any) -> set[str]:
     """Collect agent ids from assistant_id/config/context for access checks."""
     agent_ids: set[str] = set()
     if body.assistant_id and body.assistant_id not in _BUILTIN_GRAPH_IDS:
@@ -393,10 +395,99 @@ def _collect_requested_agent_ids(body: Any) -> set[str]:
         if isinstance(nested_configurable, Mapping):
             collect(nested_configurable)
 
-    return agent_ids
-
-
-async def _sync_legacy_agent_records(store) -> None:
+    return agent_ids
+
+
+def _extract_requested_model_name(body: Any) -> str | None:
+    """Best-effort read of the model override requested for this run."""
+    context = getattr(body, "context", None)
+    if isinstance(context, Mapping):
+        for key in ("model_name", "model"):
+            value = str(context.get(key) or "").strip()
+            if value:
+                return value
+
+    config = getattr(body, "config", None)
+    if isinstance(config, Mapping):
+        for container_key in ("context", "configurable"):
+            container = config.get(container_key)
+            if not isinstance(container, Mapping):
+                continue
+            for key in ("model_name", "model"):
+                value = str(container.get(key) or "").strip()
+                if value:
+                    return value
+    return None
+
+
+def _resolve_effective_run_model_name(body: Any) -> str:
+    """Resolve the model that this run will effectively use for concurrency limiting."""
+    app_config = get_app_config()
+    default_model_name = app_config.models[0].name if app_config.models else None
+    if not default_model_name:
+        raise ValueError("No chat models are configured. Please configure at least one model in config.yaml.")
+
+    requested_model_name = _extract_requested_model_name(body)
+    if requested_model_name:
+        if app_config.get_model_config(requested_model_name):
+            return requested_model_name
+        logger.warning(
+            "Requested model '%s' not found; falling back to default model '%s'.",
+            requested_model_name,
+            default_model_name,
+        )
+        return default_model_name
+
+    assistant_id = str(getattr(body, "assistant_id", "") or "").strip()
+    if assistant_id and assistant_id not in _BUILTIN_GRAPH_IDS:
+        try:
+            agent_config = load_agent_config(assistant_id)
+        except FileNotFoundError:
+            agent_config = None
+        except Exception:
+            logger.debug("Failed to load agent config for model resolution: %s", assistant_id, exc_info=True)
+            agent_config = None
+        agent_model_name = str(getattr(agent_config, "model", "") or "").strip()
+        if agent_model_name:
+            if app_config.get_model_config(agent_model_name):
+                return agent_model_name
+            logger.warning(
+                "Agent model '%s' for agent '%s' not found; falling back to default model '%s'.",
+                agent_model_name,
+                assistant_id,
+                default_model_name,
+            )
+
+    return default_model_name
+
+
+async def _enforce_model_concurrency_limit(run_mgr: RunManager, model_name: str) -> None:
+    """Reject the run when the selected model has reached its concurrent-run limit."""
+    settings = load_system_settings().model_concurrency
+    if not settings.enabled:
+        return
+
+    limit = settings.model_limits.get(model_name, settings.default_limit)
+    try:
+        limit = int(limit)
+    except (TypeError, ValueError):
+        limit = int(settings.default_limit)
+    limit = max(1, limit)
+
+    active = await run_mgr.list_active()
+    current = sum(
+        1
+        for record in active
+        if str((record.metadata or {}).get("model_name") or "").strip() == model_name
+    )
+    if current >= limit:
+        raise HTTPException(
+            status_code=429,
+            detail=f"当前模型 {model_name} 并发已达到上限（{limit}），请稍后再试。",
+        )
+
+
+async def _sync_legacy_agent_records(store) -> None:
     """Import existing folder-backed agents as built-in records.
 
     Heavy: ``list_custom_agents()`` scans ``.deer-flow/agents/`` and reads
@@ -567,19 +658,25 @@ async def start_run(
     # admin-configured mapping rather than from a per-user-visible assistant_id.
     # Those run under the "default" bucket, where a user-owned unpublished agent
     # would otherwise 404 the access check.
-    if enforce_access:
-        await _enforce_agent_access(body, request)
-    t_post_enforce = time.perf_counter()
-
-    try:
-        record = await run_mgr.create_or_reject(
-            thread_id,
-            body.assistant_id,
-            on_disconnect=disconnect,
-            metadata=body.metadata or {},
-            kwargs={"input": body.input, "config": body.config},
-            multitask_strategy=body.multitask_strategy,
-        )
+    if enforce_access:
+        await _enforce_agent_access(body, request)
+    t_post_enforce = time.perf_counter()
+
+    effective_model_name = _resolve_effective_run_model_name(body)
+    await _enforce_model_concurrency_limit(run_mgr, effective_model_name)
+
+    run_metadata = dict(body.metadata or {})
+    run_metadata["model_name"] = effective_model_name
+
+    try:
+        record = await run_mgr.create_or_reject(
+            thread_id,
+            body.assistant_id,
+            on_disconnect=disconnect,
+            metadata=run_metadata,
+            kwargs={"input": body.input, "config": body.config},
+            multitask_strategy=body.multitask_strategy,
+        )
     except ConflictError as exc:
         # 同一会话上一轮回答还没结束，又提交了新问题 → 拒绝。给用户中文提示，
         # 而不是把内部英文「Thread ... already has an active run」直接抛到前端。

offline-backend-20260512/backend/packages/harness/deerflow/config/system_settings.py

@@ -249,6 +249,26 @@ class ConcurrencyMonitorSettings(BaseModel):
     model_config = ConfigDict(extra="ignore")
 
 
+class ModelConcurrencySettings(BaseModel):
+    """Admin-configurable per-model concurrent Q&A limits."""
+
+    enabled: bool = Field(
+        default=False,
+        description="Whether per-model concurrency limiting is enabled.",
+    )
+    default_limit: int = Field(
+        default=10,
+        ge=1,
+        le=10_000,
+        description="Fallback max concurrent runs for models without an explicit override.",
+    )
+    model_limits: dict[str, int] = Field(
+        default_factory=dict,
+        description="Optional per-model overrides: model_name -> max concurrent runs.",
+    )
+    model_config = ConfigDict(extra="ignore")
+
+
 class SystemSettings(BaseModel):
     """Top-level system settings document."""
 
@@ -262,6 +282,7 @@ class SystemSettings(BaseModel):
     skill_compression: SkillCompressionSettings = Field(default_factory=SkillCompressionSettings)
     agent_modules: AgentModuleSettings = Field(default_factory=AgentModuleSettings)
     concurrency_monitor: ConcurrencyMonitorSettings = Field(default_factory=ConcurrencyMonitorSettings)
+    model_concurrency: ModelConcurrencySettings = Field(default_factory=ModelConcurrencySettings)
     model_config = ConfigDict(extra="ignore")
 
 
@@ -305,6 +326,15 @@ def get_concurrency_monitor_settings() -> ConcurrencyMonitorSettings:
         return ConcurrencyMonitorSettings()
 
 
+def get_model_concurrency_settings() -> ModelConcurrencySettings:
+    """Convenience accessor for the model-concurrency block (safe defaults on error)."""
+    try:
+        return load_system_settings().model_concurrency
+    except Exception:
+        logger.debug("Failed to read model concurrency settings; using defaults", exc_info=True)
+        return ModelConcurrencySettings()
+
+
 def save_system_settings(settings: SystemSettings) -> SystemSettings:
     """Persist the settings document atomically, returning the saved value."""
     path = _settings_path()

offline-backend-20260512/backend/tests/test_model_concurrency.py

@@ -0,0 +1,161 @@
+"""Tests for per-model concurrency limit settings and enforcement."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+from fastapi import HTTPException
+
+from app.gateway.routers import system_settings as ss_router
+from app.gateway.services import (
+    _enforce_model_concurrency_limit,
+    _resolve_effective_run_model_name,
+)
+from deerflow.config import system_settings as ss
+from deerflow.runtime.runs.manager import RunManager
+
+
+@pytest.fixture
+def tmp_runtime_home(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> Path:
+    def fake_home() -> Path:
+        return tmp_path
+
+    monkeypatch.setattr(ss, "runtime_home", fake_home)
+    return tmp_path
+
+
+def _patch_admin(monkeypatch: pytest.MonkeyPatch, user: object | None = None) -> None:
+    admin = user or SimpleNamespace(
+        id="admin1",
+        system_role="admin",
+        email="admin@example.com",
+    )
+
+    async def _optional(_request):
+        return admin
+
+    monkeypatch.setattr(ss_router, "get_optional_user_from_request", _optional)
+
+
+@pytest.mark.asyncio
+async def test_model_concurrency_settings_round_trip(
+    tmp_runtime_home: Path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    _patch_admin(monkeypatch)
+    req = SimpleNamespace()
+
+    saved = await ss_router.update_model_concurrency(
+        ss_router.ModelConcurrencyUpdateRequest(
+            enabled=True,
+            default_limit=8,
+            model_limits={"glm5": 10, "deepseek": 4},
+        ),
+        req,
+    )
+    assert saved.enabled is True
+    assert saved.default_limit == 8
+    assert saved.model_limits == {"glm5": 10, "deepseek": 4}
+
+    file_path = tmp_runtime_home / "system_settings.json"
+    payload = json.loads(file_path.read_text(encoding="utf-8"))
+    assert payload["model_concurrency"]["enabled"] is True
+    assert payload["model_concurrency"]["default_limit"] == 8
+    assert payload["model_concurrency"]["model_limits"]["glm5"] == 10
+
+    loaded = await ss_router.get_model_concurrency(req)
+    assert loaded.enabled is True
+    assert loaded.default_limit == 8
+    assert loaded.model_limits["deepseek"] == 4
+
+
+def test_resolve_effective_run_model_name_prefers_requested_model(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    app_config = SimpleNamespace(
+        models=[SimpleNamespace(name="default-model")],
+        get_model_config=lambda name: {"name": name}
+        if name in {"default-model", "glm5", "agent-model"}
+        else None,
+    )
+    monkeypatch.setattr("app.gateway.services.get_app_config", lambda: app_config)
+    monkeypatch.setattr(
+        "app.gateway.services.load_agent_config",
+        lambda _agent_id: SimpleNamespace(model="agent-model"),
+    )
+
+    body = SimpleNamespace(
+        assistant_id="custom-agent",
+        context={"model_name": "glm5"},
+        config=None,
+    )
+    assert _resolve_effective_run_model_name(body) == "glm5"
+
+
+def test_resolve_effective_run_model_name_uses_agent_default_when_request_missing(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    app_config = SimpleNamespace(
+        models=[SimpleNamespace(name="default-model")],
+        get_model_config=lambda name: {"name": name}
+        if name in {"default-model", "agent-model"}
+        else None,
+    )
+    monkeypatch.setattr("app.gateway.services.get_app_config", lambda: app_config)
+    monkeypatch.setattr(
+        "app.gateway.services.load_agent_config",
+        lambda _agent_id: SimpleNamespace(model="agent-model"),
+    )
+
+    body = SimpleNamespace(
+        assistant_id="custom-agent",
+        context={},
+        config=None,
+    )
+    assert _resolve_effective_run_model_name(body) == "agent-model"
+
+
+@pytest.mark.asyncio
+async def test_enforce_model_concurrency_limit_rejects_when_limit_reached(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(
+        "app.gateway.services.load_system_settings",
+        lambda: SimpleNamespace(
+            model_concurrency=SimpleNamespace(
+                enabled=True,
+                default_limit=5,
+                model_limits={"glm5": 2},
+            )
+        ),
+    )
+    run_mgr = RunManager()
+    await run_mgr.create("t1", metadata={"model_name": "glm5"})
+    await run_mgr.create("t2", metadata={"model_name": "glm5"})
+
+    with pytest.raises(HTTPException) as exc:
+        await _enforce_model_concurrency_limit(run_mgr, "glm5")
+    assert exc.value.status_code == 429
+    assert "并发已达到上限" in str(exc.value.detail)
+
+
+@pytest.mark.asyncio
+async def test_enforce_model_concurrency_limit_skips_when_disabled(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(
+        "app.gateway.services.load_system_settings",
+        lambda: SimpleNamespace(
+            model_concurrency=SimpleNamespace(
+                enabled=False,
+                default_limit=1,
+                model_limits={"glm5": 1},
+            )
+        ),
+    )
+    run_mgr = RunManager()
+    await run_mgr.create("t1", metadata={"model_name": "glm5"})
+    await _enforce_model_concurrency_limit(run_mgr, "glm5")